HDDM 0.8.0 documentation


Code subject responses

There are two ways to code subject responses placed in the ‘response’ column in your data file. You can either use accuracy-coding, where 1’s and 0’s correspond to correct and error trials, or you can use stimulus-coding, where 1’s and 0’s correspond to the choice (e.g. categorization of the stimulus). HDDM interprets 0 and 1 responses as lower and upper boundary responses, respectively, so in principle either of these schemes is valid.

In most cases it is more direct to use accuracy coding because the sign and magnitude of estimated drift-rate will be directly associated with performance (higher drift rate indicates greater likelihood of terminating on the accurate boundary). However, if a certain response direction or stimulus type has a higher probability of selection and you want to estimate a response bias (which could be captured by a change in starting point of the drift process; see below), you can not use accuracy coding. (For example if a subject is more likely to press the left button than the right button, but left and right responses are equally often correct, one could not capture the response bias with a starting point toward the incorrect boundary because it would imply that those trials in which the left response was correct would be associated with a bias toward the right response). Thus stimulus coding should be used in this case, using the HDDMStimCoding model. For this, add a column to your data that codes which stimulus was correct and instantiate the model like this:

model = hddm.HDDMStimCoding(data, include='z', stim_col='stim', split_param='v')

This model expects data to have a column named stim with two distinct identifiers. For identifier 1, drift-rate v will be used while for identifier 2, -v will be used. So ultimately you only estimate one drift-rate. Alternatively you can use bias z and 1-z if you set split_param='z'. See the HDDMStimCoding help doc for more information.

Include bias and inter-trial variability

Bias and inter-trial variability parameters are optional and can be included as follows:

model = hddm.HDDM(data, bias=True, include=('sv', 'st', 'sz'))


model = hddm.HDDM(data, include=('z', 'sv', 'st', 'sz'))

Where sv is inter-trial variability in drift-rate, st is inter-trial variability in non-decision time and sz is inter-trial variability in starting-point.

There is also a convenience argument that is identical to the above.

model = hddm.HDDM(data, bias=True, include='all')

Note that you can also include a subset of parameters. This is relevant because these parameters slow down sampling significantly. If a certain parameter is estimated very close to zero or fails to converge (which can happen with the sv parameter) you might want to exclude it (or only include a group-node, see below). Finally, parameter recovery studies show that it requires a lot of trials to get meaningful estimates of these parameters.

Estimate parameters for different conditions

Most psychological experiments test how different conditions (e.g. drug manipulations) affect certain parameters. You can build arbitrarily complex models using the depends_on keyword.

model = hddm.HDDM(data, depends_on={'a': 'drug', 'v': ['drug', 'difficulty']})

This will create model in which separate thresholds are estimated for each drug condition and separate drift-rates for different drug conditions and levels of difficulty.

Note that this requires the columns ‘drug’ and ‘difficulty’ to be present in your data array. For readability it is often useful to use string identifiers (e.g. drug: off/on rather than drug: 0/1).

As you can see, single or multiple columns can be supplied as values.


The presence of outliers is notoriously challenging for likelihood models, because the likelihood of a few outliers given the generative model cab be quite low. In practice, even the model we have is reasonable for a majority of trials, it may be that data from a minority of trials is not well described by this model (e.g. due to attentional lapses). HDDM 0.4 (and upwards) supports estimation of a mixture model that enables stable parameter estimation even with outliers present in the data. You can either specify a fixed probability for obtaining an outlier (e.g. 0.05 will assume 5% of the RTs are outliers) or estimate this from the data. In practice, the precise value of p_outlier does not matter. Values greater than 0.001 and less than 0.1 are sufficient to capture the outliers, and the effect on the recovered parameters is small.

To instantiate a model with a fixed probability of getting an outlier run:

m = hddm.HDDM(data, p_outlier=0.05)

HDDM assumes that outliers come from a uniform distribution with a fixed density :math:w_{outlier} (as suggested by Ratcliff and Tuerlinckx, 2002). The resulting likelihood is as follows:

\[p(RT; v, a, t) = wfpt(RT; v, a, t) * (1-p_{outlier}) + w_{outlier} * p_{outlier}\]

The default value of :math:w_{outlier} is 0.1, which is equivalent to uniform distribution from 0 to 5 seconds. However, in practice, the outlier model is applied to all RTs, even those larger than 5.

Assess model convergence

When using MCMC sampling it is critical to make sure that our chains have converged, to ensure that we are sampling from the actual posterior distribution. Unfortunately, there is no 100% fool-proof way to assess whether chains converged. However, there are various metrics in the MCMC literature to evaluate convergence problems, and if you follow some simple steps you can be more confident.

Look at MC error statistic

When calling:


There is a column called MC error. These values should not be smaller then 1% of the posterior std. However, this is a very weak statistic and by no means sufficient to assess convergence.

Geweke statistic

The Geweke statistic is a time-series approach that compares the mean and variance of segments from the beginning and end of a single chain. You can test your model by running:

from kabuki.analyze import check_geweke
print check_geweke(model)

This will print True if non of the test-statistics is larger than 2 and False otherwise. Check the PyMC documentation for more information on this test.

Visually inspect chains

The next thing to look at are the traces of the posteriors. You can plot them by calling:


This will create a figure for each parameter in your model. Here is an example of what a not-converged chain looks like:


and an example of what a converged chain looks like:


As you can see, there are striking differences. In the not-converged case, the trace in the upper left corner is very non-stationary. There are also certain periods where no jumps are performed and the chain is stuck (horizontal lines in the trace); this is due to the proposal distribution not being tuned correctly.

Secondly, the auto-correlation (lower left plot) is quite high as you can see from the long tails of the distribution. This is a further indication that the samples are not independent draws from the posterior.

Finally, the histogram (right plot) looks rather jagged in the non-converged case. This is our approximation of the marginal posterior distribution for this parameter. Generally, subject and group mean posteriors are normal distributed (see the converged case) while group variability posteriors are Gamma distributed.

Posterior predictive analysis

Another way to assess how good your model fits the data is to perform posterior predictive analysis:


This will plot the posterior predictive in blue on top of the RT histogram in red for each subject and each condition. Since we are getting a distribution rather than a single parameter in our analysis, the posterior predictive is the average likelihood evaluated over different samples from the posterior. The width of the posterior predictive in light blue corresponds to the standard deviation.

R-hat convergence statistic

Another option to assess chain convergence is to compute the R-hat (Gelman-Rubin) statistic. This requires multiple chains to be run. If all chains converged to the same stationary distribution they should be indistinguishable. The R-hat statistic compares between-chain variance to within-chain variance.

To compute the R-hat statistic in kabuki you have to run multiple copies of your model:

from kabuki.analyze import gelman_rubin

models = []
for i in range(5):
    m = hddm.HDDM(data)
    m.sample(5000, burn=200)


The output is a dictionary that provides the R-hat for each parameter:

{'a': 1.0028806196268818,
 't': 1.0100017175108695,
 'v': 1.0232548747719443}

As of HDDM 0.4.1 you can also run multiple chains in parallel. One convenient way to do this is the IPython parallel module. To launch a cluster locally, in a shell (not Python) type:

ipcluster start

This will launch the workers in the background. IPython Parallel is much more feature rich, for more information, see the IPython parallel docs.

def run_model(id):
    import hddm
    data = hddm.load_csv('mydata.csv')
    m = hddm.HDDM(data)
    m.sample(5000, burn=20, dbname='db%i'%id, db='pickle')
    return m

from IPython.parallel import Client
v = Client()[:]
jobs = v.map(run_model, range(4)) # 4 is the number of CPUs
models = jobs.get()

# Create a new model that has all traces concatenated
# of individual models.
combined_model = kabuki.utils.concat_models(models)

What to do about lack of convergence

In the simplest case you just need to run a longer chain with more burn-in and more thinning. E.g.:

model.sample(10000, burn=5000, thin=5)

This will cause the first 5000 samples to be discarded. Of the remaining 5000 samples only every 5th sample will be saved. Thus, after sampling our trace will have a length of a 1000 samples.

You might also want to find a good starting point for running your chains. This is commonly achieved by finding the maximum posterior (MAP) via optimization. Before sampling, simply call:


which will set the starting values to the MAP. Then sample as you would normally. This is a good idea in general.

If that still does not work you might want to consider simplifying your model. Certain parameters are just notoriously slow to converge; especially inter-trial variability parameters. The reason is that often individual subjects do not provide enough information to meaningfully estimate these parameters on a per-subject basis. One way around this is to not even try to estimate individual subject parameters and instead use only group nodes. This can be achieved via the group_only_nodes keyword argument:

model = hddm.HDDM(data, include=['sv', 'st'], group_only_nodes=['sv', 'st'])

The resulting model will still have subject nodes for all parameters but sv and st.

Estimate a regression model

HDDM includes a regression model that allows estimation of trial-by-trial influences of a covariate (e.g. a brain measure like fMRI) onto DDM parameters. For example, if your prediction is that activity of a particular brain area has a linear correlation with drift-rate, you could specify the following regression model (make sure to have a column with the brain activity in your data, in our example name this column BOLD):

m = hddm.models.HDDMRegressor(data, 'v ~ BOLD:C(trial_type)')

This syntax is similar to what R uses for specifying GLMs. Basically, this means that drift-rate v is distributed according to the BOLD response and trial-type. HDDMRegression will add an intercept and slope parameter for the BOLD trial-by-trial measure for each condition. The C() specifies that trial_type (which might be a data column with A and B) should be a categorical variable which will be dummy-coded.

Internally, HDDM uses Patsy for the spcification of the linear model. The `patsy documentation`_ gives a complete overview of the functionality.

You can also pass a list to linear model descriptors if you want to include multiple covariates. E.g.:

m = hddm.models.HDDMRegressor(data, ['v ~ BOLD:C(trial_type)',
                                     'a ~ Theta'])

Stimulus coding with HDDMRegression

Stimulus coding can also be implemented in HDDMRegression. The advantage of doing so is that more complex designs, including within participant designs can be analysed (see below). The disadvantage is the HDDMRegression is slower than HDDMStimCoding.

To implement stimulus coding for z one has to define a link function for HDDMRegression:

import hddm
import numpy as np
from patsy import dmatrix

def z_link_func(x, data=mydata):
    stim = (np.asarray(dmatrix('0 + C(s,[[1],[-1]])', {'s':data.stimulus.ix[x.index]})))
    return 1 / (1 + np.exp(-(x * stim)))

Similarly, the link function for v is:

def v_link_func(x, data=mydata):
    stim = (np.asarray(dmatrix('0 + C(s,[[1],[-1]])', {'s':data.stimulus.ix[x.index]})))
    return x * stim

To specify a complete model you have to define a complete regression model and submit it with the data to the model.

z_reg = {'model': 'z ~ 1 + C(condition)', 'link_func': z_link_func}
v_reg = {'model': 'v ~ 1 + C(condition)', 'link_func': lambda x: x}
reg_model = [z_reg, v_reg]
hddm_regrssion_model = hddm.HDDMRegressor(data, reg_model, include='z')

Of course, your model could also regress either z or v. For example

v_reg =  [{'model': 'v ~ 1 + C(condition)', 'link_func': v_link_func, group_only_regressors=True}]
hddm_regrssion_model = hddm.HDDMRegressor(data, v_reg, include='z')

For a more elaborate example and parameter recovery study using HDDMRegression, see the tutorial on using HDDMRegression for stimulus coding.

Perform model comparison

We can often come up with different viable hypotheses about which parameters might be influenced by our experimental conditions. Above you can see how you can create these different models using the depends_on keyword.


To compare which model does a better job at explaining the data you can compare the DIC scores (lower is better) emitted when calling:


DIC, however, is far from being a perfect measure. So it shouldn’t be your only weapon in deciding which model is best.

Posterior predictive check

A very elegant method to compare models is to sample new data sets from the estimated model and see how well these simulated data sets corresponds to the actual data on some measurement (e.g. is the mean RT well recovered by this model?).

The best place to learn about this is :ref:Posterior Predictive Tutorial <chap_tutorial_post_pred>.

Run Quantile Opimization

Even though Hierarchical Bayesian estimation tends to produce better fit – especially with few number of trials – it is quite a bit slower than the Quantile optimization method that e.g. Roger Ratcliff uses. If you have lots of data (>100 trials per condition) and you don’t care about posterior estimates you can use the HDDM.optimize() method to run quantile optimization. The first argument is a string identifier of which optimization method you want to run and can be one of chisquare, gsquare or ML.

model = hddm.HDDM(data, depends_on={'v': ['word_freq', 'reps']})
params = model.optimize('chisquare')

Note that this function will by default not estimate individual subject parameters but rather do quantile averaging to only estimate group parameters. Running different models for each individual subject is quite easy however:

subj_params = []
for subj_idx, subj_data in data.groupby('subj_idx'):
   m_subj = hddm.HDDM(subj_data, depends_on={'v': ['word_freq', 'reps']})
params = pandas.DataFrame(subj_params)

Save and load models

HDDM models can be saved and reloaded in a separate python session. Note that you have to save the traces to file by using the db backend.

model = hddm.HDDM(data, bias=True)  # a very simple model...
model.sample(5000, burn=20, dbname='traces.db', db='pickle')

Now assume that you start a new python session, after the chain started above is completed.

model = hddm.load('mymodel')

HDDM uses the pickle module to save and load models.

Compare parameters to other papers

A lot of people are very used to the parameters that come from Ratcliff’s assumption, which is that the noise coefficient is 0.1, rather than 1. That noise parameter is a “scaling parameter”, meaning that you just need to multiply HDDM’s estimates of drift, starting point, and threshold (as well as variabilities in these parameters) by 0.1, and you get estimates that are commensurate with most of the parameter estimates in the literature. Ratcliff et al. mention that you can scale all the DDM parameters in that way and you thereby get identical choice probabilities and RTs in several papers, such as Ratcliff & Tuerlinckx, 2002.

Moreover, the z parameter in HDDM is relative to a (i.e. it ranges from 0 to 1). Ratcliff usually reports z as an absolute value (i.e. it ranges from 0 to a). To transform HDDM z to Ratcliff’s notation use a*z.

Hypothesis testing

Since HDDM uses Bayesian estimation it is straight forward to analyze the posterior directly for hypothesis testing. You can directly work with the posterior and ask statistically meaningful questions. For example, assume you two conditions for drift-rate:

v_Win, v_Neutral= m.nodes_db.node[['v(Win)', 'v(Neutral)']]
print "P_v(Win > Neutral) = ", (v_Win.trace() > v_Neutral.trace()).mean()

Would give you the probability that the drift-rate in the Win condition is higher than in the Neutral condition.

Note that it is wrong to just input the subject parameters of a hierarchical into a frequentist test like the T-test. The hierarchical model violates the independence assumption.

There are some excellent books on Bayesian data analysis for cognitive science, e.g. Lee & Wagenmakers and Kruschke, or see the BEST paper for a single journal article comparing Bayesian estimation to the t-test.